Inventory QC

OpenCV 5 + AWS system that detects blurry inventory files and triggers actions in a perception-decision-action loop.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Inventory QC

Inventory QC OpenCV 5 + AWS system that detects blurry inventory files and triggers actions in a perception-decision-action loop.

Inspiration

Warehouse and retail workers waste hours sorting blurry, unusable inventory photos. Photos are taken quickly, sometimes in poor lighting or with shaky hands, and the low-quality ones slip through — causing inventory errors, returns, and lost productivity.

Existing tools either don't check quality at all, or require expensive manual review. I wanted to build something that automatically catches bad photos at the source, before they enter the inventory system, using only serverless AWS services and open-source computer vision.

The OpenCV AI Competition 2026 theme of Agentic Vision — where visual understanding drives real decisions and actions — was the perfect fit.

What it does

Inventory QC is an automated quality control system for inventory images and videos. Users upload a file through a web interface (or drop it directly into S3). The system:

  1. Perceives — OpenCV 5 computes the sharpness score using Laplacian variance. For videos, it extracts up to 10 frames and averages the scores.
  2. Decides — The agent compares the score to a threshold (100). If below, the file is rejected. If above, it's accepted.
  3. Acts — If rejected, an AWS SNS notification is sent to the user to retake the photo. If accepted, no action is needed.

This is a real perception -> decision -> action loop, exactly what the Agentic Vision track asks for.

How I built it

Architecture

User (web interface) -> API Gateway -> AWS Lambda (OpenCV 5) -> Amazon S3 (image storage) -> Amazon SNS (rejection alerts). Web interface is hosted on S3 static hosting.

The same Lambda function handles two event types:

  • S3 events (when a file is dropped directly into the bucket)
  • API Gateway events (when the user uploads via the web interface)

OpenCV 5 Implementation

The core of the vision analysis is Laplacian variance — a well-known, computationally cheap method for blur detection:

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) score = cv2.Laplacian(gray, cv2.CV_64F).var()

A sharp image has high variance (lots of high-frequency edges); a blurry image has low variance. The threshold was calibrated empirically:

floue.jpg (image): score 6.55 -> REJECTED nette.jpg (image): score 6059.05 -> ACCEPTED test_video.mp4 (video, 10 frames): score 41949.01 -> ACCEPTED

The gap between 6.55 and 6059.05 shows the method is highly discriminative.

AWS Deployment

  • OpenCV 5 was packaged as a Lambda Layer (77 MB) using a Linux-compatible wheel (manylinux2014_x86_64).
  • Lambda runs Python 3.13 with 1 GB RAM and a 30-second timeout.
  • API Gateway exposes a POST endpoint /analyze with Lambda Proxy integration.
  • S3 stores uploads and hosts the static web interface.
  • SNS sends email notifications on rejection.

Web Interface

A cyberpunk HUD-style interface built in pure HTML/CSS/JavaScript — no frameworks. Features:

  • Animated boot screen with a camera crosshair logo and system log
  • Military crosshair reticle background
  • Particle cloud that reacts to mouse movement
  • Real-time analysis with clear accept/reject feedback
  • Handles both images and videos — a video preview for uploads, img for images
  • Reset button to analyze another file without reloading

Challenges I faced

  1. Packaging OpenCV 5 for Lambda — The default pip install opencv-python produces a Windows binary that won't run on Lambda's Linux environment. I had to use pip install opencv-python-headless --platform manylinux2014_x86_64 and package the result as a Lambda Layer with the correct python/ directory structure.

  2. CORS on API Gateway — The browser blocked requests from the S3-hosted interface to the API Gateway domain. I had to enable CORS on the resource and ensure the Lambda returned Access-Control-Allow-Origin: * headers.

  3. Two event types in one Lambda — S3 and API Gateway send completely different event structures. I built a dispatcher that detects the event type and adapts the parsing logic.

  4. Video size limits — API Gateway has a 10 MB payload limit. Base64 encoding inflates the size by 33%, so videos must stay under ~7 MB. This is a documented limitation.

  5. Timeout and memory — The default Lambda settings (3s, 128 MB) were far too small for OpenCV. I bumped them to 30s and 1 GB.

What I learned

  • How to package native libraries (OpenCV) for AWS Lambda
  • How to build a perception-decision-action loop with serverless services
  • The importance of CORS in distributed web applications
  • How to handle multiple event sources with a single Lambda function
  • Video analysis: extracting frames with cv2.VideoCapture and averaging scores

What's next

  • Real-time live mode using WebSocket API Gateway
  • Object detection with YOLO (e.g., detecting missing items in inventory photos)
  • Multi-file batch processing
  • COOL (Cloud-Optimized OpenCV) on AWS Graviton for lower latency and cost — a natural extension toward the Best Use of COOL track

Why this matters

Manual inventory photo review costs businesses time and money. Inventory QC automates the first line of defense — catching bad files before they enter the system. Because it's entirely serverless, it scales to thousands of uploads per day at near-zero cost, and it can be deployed in any AWS region in minutes.

The system is a complete, working proof-of-concept that demonstrates how OpenCV 5 and AWS can combine to solve a real-world problem — and it's ready to grow into a production-grade tool. https://youtu.be/GOtx6mhf_CU https://github.com/vladimirjoseph36-png/inventory-qc http://inventory-qc-website-vladi-2026.s3-website-us-east-1.amazonaws.com OpenCV 5, Python 3.13, AWS Lambda, Amazon S3, Amazon SNS, Amazon API Gateway, HTML, CSS, JavaScript, AWS IAM Agentic Vision Vladimir Joseph Inventory QC uses OpenCV 5 Laplacian variance analysis on AWS Lambda to automatically reject blurry inventory images and videos, triggering SNS notifications in a perception-decision-action loop.

Built With

Share this project:

Updates

Submission history